Re: [PATCH v3] vsprintf: Prevent crash when dereferencing invalid pointers

[Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>]

On (03/15/18 18:35), Linus Torvalds wrote:
> On Thu, Mar 15, 2018 at 6:18 PM, Sergey Senozhatsky
> <sergey.senozhatsky.work@gmail.com> wrote:
> >
> > Hm, may be sizeof(ptr) still won't suffice. It would be great if we
> > could always look at spec.field_width, which can be up to 2 * sizeof(void *),
> > and then just probe_kernel_read(spec.field_width). E.g., %b/%bl prints out a
> > bitmap, accessing max_t(int, spec.field_width, 0) bits, which is good. But,
> > for instance, %U (uuid printout) doesn't look at spec.field_width, and reads
> > in 16 bytes from the given memory address. Then we have ipv4/ipv6, mac, etc.
> > So I think that checking just 1 byte or sizeof(ptr) is not really enough if
> > we want to fix vsprintf. What do you think?
> 
> Honestly, I think it would be better to move the whole logic to the
> functions that actually do the printout.
> 
> Then you can do it right, and you don't need to have the strchr() either.
>
> There really isn't any commonality between the different versions.
> field_width is meaningless, since it's about the size of the _printed_
> field, not the size in memory.

Agreed!

> Would it be a few more lines? Yes. But it would also clarify the code
> and get all the cases right. Look at hex_string() for example, and
> imagine fetching a byte at a time and just getting the corner cases
> automatically right.

So, basically, what I tried to say - any byte past the first sizeof(ptr)
bytes or past the first byte that we check_access() can cause problems,
which this patch is trying to address. As an example, FORMAT_TYPE_STR
case

	printk("%.*s\n", p->buf)
	 vsnprintf()
	  string()

Where ->buf is a _nearly always_ properly nul terminated char buf[128]
array in struct foo. So moving that check_access() to every function that
does printout sounds good to me, as well as checking every byte we access
[assuming that we want to cure vsprintf], not just the first one or the
first sizeof(ptr) bytes.

	-ss